Download IJMC Feature selection and classification of microarray gene

Iranian Journal of Mathematical Chemistry, Vol. 4, No.2, December 2013, pp. 163 175
IJMC
Feature selection and classification of microarray
gene expression data of ovarian carcinoma patients
using weighted voting support vector machine
S. MASOUM AND S. GHAHERI
Department of Analytical Chemistry, Faculty of Chemistry, University of Kashan, I. R. Iran
(Received October 1, 2012; Accepted Feb 18, 2013)
ABSTRACT
We can reach by DNA microarray gene expression to such wealth of information with
thousands of variables (genes). Analysis of this information can show genetic reasons of
disease and tumor differences. In this study we try to reduce high-dimensional data by
statistical method to select valuable genes with high impact as biomarkers and then classify
ovarian tumor based on gene expression data of two patient groups. One group treated by
standard therapies and survived, while another group didn’t be cure and die after some times.
In the first step we used weighted voting algorithm (WVA) for selecting impressive genes to
reduce dimension, therefore eliminate noisy data and make analysis easier and then partial
least square – discriminante analysis (PLS-DA) and support vector machine (SVM) methods
have been applied for classification of diminished data. Results show that classification by
PLS-DA can distinguish two groups somewhat but SVM is more efficient and sufficient
classification method.
Keywords: weighted voting algorithm, support vector machine, tumor classification, ovarian
cancer, gene expression data.
1. INTRODUCTION
In the past decades, chemometrics methods generally have been used to solve chemical
problems. But today, there is an approach for using these methods to analysis of gene
expression data [1-3]. DNA microarrays are capable of detecting the expression levels of
thousand genes over a few tens of different samples simultaneously [4]. Because of such
huge volume of data, there is an increasing attention in data mining field and extraction of
precious and helpful information from a huge collection of data [5]. Using statistical
method and data mining is necessary to understand the mechanism and process of human
deceases. Human demise as cancer and tumor are because of gene expression changes, so

Corresponding author (Email: [email protected])
164
S. MASOUM AND S. GHAHERI
discrimination between different group of samples or between patients and healthy people
based on their gene expression is important step to take information out. By classification
of individual gene expression, we have a model that can predict the class of a goal sample
with unknown class label. The enormous amount of data and small size of individuals is a
challenge of this kind of studies, because in such immense data, arrangement is
complicated. Also this kind of data usually contains unsuitable and redundancy information
[6]. Furthermore, before classification, reduce of dimension is important task. By
decreasing dimension and select some genes through whole, we can use selected genes as
biomarker to forecast deceases. Researchers introduce many methods for classification and
selecting best features. For example, Sarhan developed a method based on artificial neutral
network (ANN) and discrete cosine transforms (DCT) [7]. Zheng et al. proposed
independent component analysis (ICA) method coupled by sequential floating forward
selection (SFFS) technique [8]. Literatures also referred to PLS-DA frequently [9-11].
Furey et al. offered SVM method for classification of gene expression [2]. Other methods
that have been used are radial basis function neural network [12] logistic discrimination and
quadratic discriminante analysis [13].
Ovarian carcinoma is one of the most common type of gynecological cancers, is
fifth reason of cancer demise in women [14]. Standard treatment in ovarian cancer patient
is surgery followed by chemotherapy that some patient will be cured while others relapse.
If these different patient groups could be identified before therapy, the alternative
treatments or strategies might be used instead of standard treatments [15].
The data set have been used in this study is microarray analysis results of ovarian
cancer patients. So, studied samples are consisting of gene expressions of two groups. One
group didn’t be cure and die after some time and another one are survived after 5 years. All
samples are belonging to patients in stage ш ovarian adecarcinoma. We used WVA to
reduce dimension and then classify reduced data by PLS-DA and SVM methods.
2. METHODOLOGY
1.1. Weighted voting algorithm (WVA)
The first step in data analysis is valuation of genes as variables. Selecting genes with higher
and lower expression is important task. Using these genes as prognostic factors may make
easy identification of patients who are expected to relapse and die of the decease [15].
Furthermore, because of the large number of variables, recourse to conventional
classification methods may be hard both for analytical and interpretive reasons. In this
study we used WVA to select genes with higher and lower expression as biomarkers [1617]. This algorithm calculates Sx value for each genes of data set according to equation 1.
Feature selection and classification of microarray gene expression data
Sx =
(  c1   c 2 )
( c1   c 2 )
165
(1)
Sx= weighted voting value for every gene
µc = mean of expression values in class 1 and class 2
σc = standard deviation of expression values in class 1 and class 2
The Sx value show how much is correlation of every genes with a particular
distinction. Also it detects genes which have higher variance in one group but low variance
in another one. This bias is useful for biological sample. For example, in cancer research,
genes in normal tissue work normally and the regulation of which are strict. However, in
tumors, genes are deregulated and levels of microarray data expressions vary widely [18].
After dimension reduction, the selected genes were used for classification.
1.2. Partial Least Square – Discriminant Analysis (PLS-DA)
PLS–DA, a special form of partial least square (PLS) modeling, aims to find the variables
and direction in multivariate space, which discriminate the known class in training set. In
PLS-DA, an indicator Y matrix of category variables is constructed which contains as many
columns as there are known class in the training set. In this context, PLS-DA accomplishes
a rotation of the projection to latent variable focusing on class separation [19]. PLS-DA
score plot show distribution of two classes and root mean square error (RMSE) value
reveals validity of separation. RMSE as prediction error parameter defined as below:
RMSE =
i (i  ˆ i ) 2
n
(2)
i = real class for ith sample
̂i = predicted class for ith sample
n= number of sample
It is obvious that the best parameter for RMSE is 0 when model can predict all
classes exactly right.
3.2. Support Vector Machine (SVM)
The SVM algorithm originally introduced by V. Vapkin in 1998 [20]. For the first time,
Furey et al. offered SVM method for microarray expression classification [2]. SVM
classification is based on hyper-plan or a set of hyper-plans that separate labeled training
data considering their classes so that the distance between them will be maximized. If in a
166
S. MASOUM AND S. GHAHERI
finite dimension space (linear), separation isn’t possible, a much higher dimension on
infinite space is used in combination with kernel techniques such as linear, polynomial,
Gaussian radial basis and exponential radial basis function. It’s clear that the hyper-plan
which can classify the two classes of samples suitably isn’t unique. To finding the best
hyper-plan, called optimal separating hyperplan (OSH), the concepts of margin, is
introduced as distance of hyper-plan to nearest data point of each class (support vectors).
There are many classifiers called hyper-plan that can separate the data, but there is only one
that maximizes the margin [21]. Suppose problem of separating the set of training vectors
belonging to two separate classes,
D = {(x1, y1), …,( xm ,ym)}, x Rn , y (1,-1)
(3)
The hyperplan is:
w, x + b = 0
(4)
Where w is the normalized weight vector with the same dimension as x and b is the
normalized bias of the hyper-plane, any hyper-plane f(x) should meet the following state:
w, x + b  1 for yi =1
(5)
w, x + b  -1 for yi = -1
(6)
So:
yi (w, x + b )  1
(7)
Then, the margin between the two paralleled hyper-planes can be written as:
2
Margin 
(8)
w
Therefore, the structure of OSH can be transformed to the following optimizing problem:
2
Maximize:
w
Subject to: w ,x + b 1
By solving problem and finding OSH, we can classify a new data sample s. A label
is assigned in according to its relationship to the decision boundary, and the corresponding
decision function is:
f(s) = sign (w,s + b)
(9)
4.2. Dataset
The data base resource currently available on the World Wide Web: www.ncbi.nih.gov/geo
was a table (X-matrix) in which 56 individuals with 30000 probe sets (variables) reported.
Each probe set contains one gene and it is also possible that one gene occupies more than
one probe set. Individuals are belonging to two groups, 5year survivor (class 1) and dead
(class 2) patients. If consider X as a descriptor matrix, an appropriately selected dependant
Feature selection and classification of microarray gene expression data
167
variable matrix (the dummy matrix Y) designating membership to given class. The data set
was divided into two sets of training (38 samples) and monitoring (16 samples). The
training set was used to develop the model. Together with the performance of the training
set, the performance of an independent set must also monitored (monitoring set) to obstruct
the overtraining phenomena.
All computations and chemometrics analyses were executed with programs in
Matlab v. 7. (The mathworks, Inc., Natick, MA, USA). Different algorithm has been
proposed in the literature to perform SVM for classification [22-24]. The Lin’s Lib SVM v.
2.33 algorithm was used in the present work [24].
2. RESULTS AND DISCUSSION
The weighted voting algorithm makes a weighted linear combination of relevant marker or
informative feature obtained in the data set. Fifty probe sets with lowest and fifty probe sets
with highest value of Sx were selected as biomarkers are listed in table 1 and 2,
respectively. To evaluate the robustness of these biomarkers, the final step is to classify the
data set. There have been many methods for performing the classification task. We used
PLS-DA and SVM which have been proved to be very useful and robust to classify the
microarray gene expression data.
Modeling by PLS-DA method was done on diminished training set. Validation of
model was checked by monitoring set. Different preprocessing may help better
discrimination. RMSE values with different preprocessing methods are arranged in table 3.
In Figure 1, the result for the three latent variable normalize-PLS-DA model is shown. This
figure shows that PLS-DA method can separate two groups somewhat but not completely.
PLS-DA result based on original data set without any dimensional reduction in table 4
indicates that RMSE value for monitoring set is not satisfactory compare to reduce one.
Among different supervised methods, SVM seems to be the most suitable one, because for
the classification purpose only support vectors are needed. This means that for the
classification a limited number of data points are used and therefore the calculation
processes would be reduced. In the present work, among 38 samples of the training set only
a total of 14 samples were chosen as support vectors. When it is used for classification,
SVM can separate a given set of binary labeled training data with a hyper-plan that is
maximally distant from them (the maximal margin hyper-plan). For the case in which no
linear separation is possible, they can work in combination with the technique of kernels,
which automatically realize a nonlinear mapping to a feature space. Generally, the hyperplan founded by SVM in a feature space corresponds to a nonlinear decision boundary in
the original space. Linear SVM results show the 100% accuracy on training and 93%
accuracy on monitoring set. Applying WVA-SVM on DNA microarray data can be
considered as a powerful tool for tumor classification from gene expression data.
168
S. MASOUM AND S. GHAHERI
Previous study on such data set show that three genes are candidate biomarkers:
TACC1 (transforming acidic coiled-coil containing protein 1), MUC5B (mucin 5 subtype
B) and PRAME (preferentially expressed antigen in melanoma) [15]. The typical function
of TACC1 is not accurately known, but observations have shown that the protein is
concentrated at centrosomes during mitosis and may play a role in cytokinesis [25,26]. This
gene known as a cancer related feature in literature [27-29]. MUC5B belongs to the mucin
family of high-molecular-weight glycoproteins found in human epithelial cells. MUC5B, a
secreted gel forming mucin, has been studied and play important role in a number of tumor
types like breast and gastric cancer [30,31]. The function of PRAME in normal tissue is
still unknown, but it encodes an antigen recognized by autologous cytolytic T lymphocytes
and its expression is absent or low in normal adult tissue, except male germ cells [32].The
effect of this gene, as cancer related gene in ovarian cancer has been verified in some
researches [34,35]. Also abnormal expression of this gene is observed in melanoma and
neuroblastoma cancer [35,36]. Results show Sx values for these three genes obtained by
weighted voting algorithm are approximately in good agreement with other studies. When
gene expression of survivor subgroup compared with remaining tumor, TACC1 and
MUC5B are between highest Sx and PRAME is one of fifty genes with lowest Sx. Various
genes with unknown function among the “top 100” (50 in table 1 and 50 in table 2) deserve
high priority in future studies, that provide shortcuts in genome-based ovarian cancer
research.
3. CONCLUSION
In this research, we presented WVA and SVM for feature selection and classification of
tumor, based on microarray gene expression data. The methodological involve dimension
reduction of high-dimensional gene expression data, followed by feature selection using
WVA and classification by applying SVM. The results show that our method is effective
and efficient in classifying ovarian tumor.
ACKNOWLEDGMENT
The authors are grateful to University of Kashan for supporting this work by Grant NO.
159181/3.
Feature selection and classification of microarray gene expression data
169
Table 1: Fifty Probe Sets with Lowest Sx.
No.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
Identifier
ALPL
TSPYL5
AK023883
Operon oligo ID:
300001540
C10orf46
GPR137C
ZNF250
HSD17B14
TMTC1
GALNT2
NASP
XM_499130
NRBP1
ASL
TMTC1
EFNA4
COL17A1
C19orf62
RNF185
FAM84A
DNMT3A
GPAA1
RBMY1J
ATP2B4
C20orf117
Sx
-0.8387509
-0.7305065
-0.7115028
No.
26
27
28
Identifier
EGFL6
C22orf28
LUC7L3
Sx
-0.5806355
-0.5791292
-0.5784992
-0.6912204
29
PRAME
-0.5778165
-0.6911821
-0.6905153
-0.688837
-0.6707914
-0.6657545
-0.6639226
-0.6602168
-0.6564458
-0.6536662
-0.6490638
-0.6417303
-0.6353448
-0.6033936
-0.6029183
-0.599269
-0.5991216
-0.5969614
-0.5966022
-0.5936286
-0.5850159
-0.5840006
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
KLHL24
RPN2
ARID4B
CST3
MAPK8IP1
HM13
RBM42
AK025101
HIST2H4A
SERTAD3
SEC61A1
FLJ21369
ADAM17
CGN
SNRPB
COPB2
ERH
IFIT1
NOTCH3
C20orf117
KRT6A
-0.5755147
-0.5733304
-0.5729077
-0.5720629
-0.5713276
-0.5698085
-0.5633995
-0.5625916
-0.5591982
-0.5591699
-0.5590856
-0.555698
-0.5539212
-0.5528546
-0.5528253
-0.5525224
-0.5511848
-0.5480109
-0.5475582
-0.5474255
-0.544361
170
S. MASOUM AND S. GHAHERI
Table 2: Fifty Probe Sets with Highest Sx.
No. Identifier
1
INTS10
2
XM_036708
Operon oligo
3
ID: 300002269
4
CDH3
5
TACC1
6
SPATA2L
Operon oligo
7
ID: 200020491
8
APOH
9
FBXL21
10 GBE1
11 MLC1
12 PSD
Sx
0.77754612
0.6892122
No. Identifier
26 SLC1A1
27 KIAA0415
Sx
0.55741585
0.55542346
0.68419241
28
BC012900
0.55444906
0.67859078
0.65699497
0.65636689
29
30
31
POLR3D
TTC18
HIC2
0.55138735
0.55021117
0.54892664
0.6374128
32
CAPS
0.54833025
0.63583484
0.63469991
0.63243051
0.61475509
0.611211
33
34
35
36
37
0.54677712
0.54652939
0.54024783
0.53942265
0.53650802
13
CLU
0.61094734
38
14
15
16
17
18
19
20
21
XM_379145
WDR46
SAMD11
MUC5B
PHLDA1
KIAA1462
RELN
MRM1
Operon oligo
ID: 200006958
XM_496984
TSEN2
NM_031306
0.60190264
0.58562923
0.580747
0.57932012
0.57808495
0.57554463
0.56497197
0.56484086
39
40
41
42
43
44
45
46
ABHD14B
DNAH9
HCN2
XM_496691
NUDT6
Operon oligo
ID: 300002652
ANKRD18A
FYN
FAM174A
UCN
GIPC3
C7orf34
CHD9
NP_689472
0.56451803
47
LCAT
0.5280964
0.56330992
0.55978275
0.55795636
48
49
50
C16orf45
B9D1
KLHDC7A
0.5277279
0.52739697
0.52606313
22
23
24
25
0.53470338
0.53420268
0.53414623
0.53145332
0.53117405
0.53083502
0.53059775
0.52966783
0.52822182
Feature selection and classification of microarray gene expression data
171
Table 3: PLS-DA results with different preprocessing on reduced data.
Noun
Standard Normal
Variate (SNV)
Orthogonal Signal
Correction (OSC)
Normalize
class 1
0.15
class 2
0.15
class 1
0.35
Class 2
0.37
0.08
0.12
0.32
0.34
0.13
0.10
0.35
0.39
0.10
0.12
0.30
0.32
Table 4: PLSDA Results with Different Preprocessing on Original Data.
Noun
Standard Normal
Variate (SNV)
Orthogonal Signal
Correction (OSC)
Normalize
class 1
0.12
class 2
0.11
class 1
0.51
Class 2
0.56
0.06
0.06
0.43
0.47
0.05
0.05
0.43
0.46
0.07
0.08
0.47
0.53
172
S. MASOUM AND S. GHAHERI
Figure.1: Score plot of the first three latent vectors. Training set: 5-year survivor (), dead
(), corresponding monitoring set samples (,).
REFERENCES
1. U. Alon and M. Barkai, Broad patterns of gene expression revealed by clustering
analysis of tumor and normal colon tissues probed by oligonucleotide arrays, P.
Natl. Acad. Sci. USA. 96 (1999) 67456750.
2. T.S. Furey, N. Cristianini, N. Duffy, D.W. Bednarski, M. Schummer and D.
Haussler, Support vector machine classification and validation of cancer tissue
samples using microarray expression data, Bioinforma, 16 (2000) 906914.
Feature selection and classification of microarray gene expression data
173
3. M. Bittner, P. Meltzer, Y. Chen, Y. Jiang, E. Seftor, M. Hendrix, M. Radmacher, R.
Simon, Z. Yakhinik and N. Sampask, Molecular classi fcation of cutaneous
malignant melanoma by gene expression profling, Nature. 406 (2000) 536540.
4. C.H. Zheng, Y.W. Chong and H.Q. Wang, Gene selection using independent
variable group analysis for tumor classification, Neural. Comput. Appl. 20 (2011)
161–170.
5. J.P. Bigus, Data mining with neural networks: solving business problems from
application development to decision support, (McGraw-Hill, Hightstown New
Jersey, 1996).
6. A. Osareh and B. Shadgar, Classification and diagnostic prediction of cancers using
gene microarray data analysis, J. Appl. Sci. 9 (2009) 452458.
7. A. Sarhan, Cancer Classification Based on Micro array Gene Expression Data
Using DCT and ANN, J. Theor. Appl. Inf. Tech. 6 (2009) 208–216.
8. C. Zheng, D. Huang and L. Shang, Feature selection in independent component
subspace for microarray data classification, Neurocomputing. 69 (2006)
24072410.
9. G. Musumarra, V. Barresi, D.F. Condorelli and S. Scir , A bioinformatic approach
to the identification of candidate genes for the development of new cancer
diagnostics, Biol. Chem. 384 (2003) 321–327.
10. M. Pérez-Enciso and M. Tenenhaus, Prediction of clinical outcome with microarray
data: A partial least squares discriminant analysis (PLS-DA) approach, Hum. Genet.
112 (2003) 581592.
11. M. Barker and W. Rayens, Partial least squares for discrimination, J. Chemometr.
17 (2003) 166173.
12. A. Castaño, F. Fernández-Navarro, C. Hervás-Martínez and P.A. Gutierrez, Neurologistic Models Based on Evolutionary Generalized Radial Basis Function for the
Microarray Gene Expression Classification Problem, Neural. Process. Lett. 34
(2011) 117-131.
13. S. Dudoit, J. Fridlyand and T.P. Speed, Comparison of discrimination methods for
the classification of tumors using gene expression data, J. Am. Stat. Assoc. 97
(2002) 77–87.
14. I.B. Runnebaum and E. Stickeler, Epidemiological and molecular aspects of ovarian
cancer risk, J. Cancer. Res. Clin. 127 (2001) 7379.
15. K. Partheen, K. Levan, L. Osterberg and G. Horvath, Expression analysis of stage
III serous ovarian adenocarcinoma distinguishes a sub-group of survivors, Eur. J.
Cancer, 42 (2006) 28462854.
16. S. Ramaswamy, K.N. Ross, E.S. Lander and T.R. Golub, A molecular signature of
metastasis in primary solid tumors, Nat. Genet. 33 (2002) 49–54.
174
S. MASOUM AND S. GHAHERI
17. T. J. MacDonald, K.M. Brown, B. LaFleur, K. Peterson, C. Lawlor, Y. Chen, R.J.
Packer, P. Cogen and D. A Stephan, Expression profiling of medulloblastoma:
PDGFRA and the RAS/MAPK pathway as therapeutic targets for metastatic
disease, Nat. Genet. 29 (2001) 143152.
18. M. Reich, K. Ohm, M. Angelo, P. Tamayo and J.P. Mesirov, GeneCluster 2.0: an
advanced toolset for bioarray analysis, Bioinforma. 20 (2004) 17971798.
19. K.P. Singh, A. Malik, D. Mohan, S. Sinha and V.K. Singh, Chemometric data
analysis of pollutants in wastewater–a case study, Anal. Chim. Acta. 532 (2005)
15–25.
20. V. Vapnik, the nature of statistical learning theory, (Springer, New york, 1998).
21. http://www.eas.uccs.edu/wang/ECE5990/SVM.pdf, S R. Gunn, 1998.
22. http://www.csie.ntu.edu.tw/~cjlin/libsvm, C.-C. Chang and C.-J. Lin, 2002.
23. H. Li, Y. Liang and Q. Xu, Support vector machines and its applications in
chemistry, Chemometr. Intell. Lab. 95 (2008) 188198.
24. http://www.kernel-machines.org/
25. F. Gergely, C. Karlsson, I. Still, J. Cowell, J. Kilmartin and J.W. Raff, The TACC
domain identifies a family of centrosomal proteins that can interact with
microtubules, P. Natl. Acad. of Sci. USA. 97 (2000) 1435214357.
26. B. Delaval, A. Ferrand, N. Conte, C. Larroque, D. Hernandez-Verdun, C. Prigent
and D. Birnbaum, Aurora B -TACC1 protein complex in cytokinesis, Oncogene. 23
(2004) 45164522.
27. L.W. Chu, P. Troncoso, D. a Johnston and J.C. Liang, Genetic markers useful for
distinguishing between organ-confined and locally advanced prostate cancer, Gene.
Chromosome. Canc. 36 (2003) 303312.
28. K. Partheen, K. Levan, L. Osterberg, K. Helou, G. Horvath, Analysis of cytogenetic
alterations in stage III serous ovarian adenocarcinoma reveals a heterogeneous
group regarding survival, surgical outcome, and substage, Gene. Chromosome.
Canc. 40 (2004) 342348.
29. R. Anbazhagan, H. Fujii and E. Gabrielson, Allelic loss of chromosomal arm 8p in
breast cancer progression, Am. J. Pathol. 152 (1998) 815819.
30. N. Berois, M. Varangot, C. Sóñora, L. Zarantonelli, C. Pressa, R. Laviña, J.L.
Rodríguez, F. Delgado, N. Porchet, J. P. Aubert and E. Osinaga, Detection of bone
marrow-disseminated breast cancer cells using an RT-PCR assay of MUC5B
mRNA, Int. J. Cancer. 103 (2003) 550555.
31. M. Perrais, P. Pigny, M.P. Buisine, N. Porchet, J.P. Aubert and I. Van SeuningenLempire, Aberrant expression of human mucin gene MUC5B in gastric carcinoma
and cancer cells. Identification and regulation of a distal promoter, J. Biol. Chem.
276 (2001) 1538615396.
Feature selection and classification of microarray gene expression data
175
32. H. Ikeda, B. Lethé, F. Lehmann, N. van Baren, J.F. Baurain, C. de Smet, H.
Chambost, M. Vitale, A. Moretta, T. Boon and P.G. Coulie, Characterization of an
antigen that is recognized on a melanoma showing partial HLA loss by CTL
expressing an NK inhibitory receptor, Immunity. 6 (1997) 199208.
33. T.R. Adib, S. Henderson, C. Perrett, D. Hewitt, D. Bourmpoulia, J. Ledermann, C.
Boshoff, Predicting biomarkers for ovarian cancer using gene-expression
microarrays, Brit. J. Cancer. 90 (2004) 686692.
34. K. Hibbs, K.M. Skubitz, S.E. Pambuccian, R.C. Casey, K.M. Burleson, T.R.
Oegema, J.J. Thiele, S.M. Grindle, R.L. Bliss and A.P.N. Skubitz, Differential gene
expression in ovarian carcinoma: identification of potential biomarkers, Am. J.
Pathol.165 (2004) 397414.
35. K.H. Lu, A.P. Patterson, L. Wang, R.T. Marquez, E.N. Atkinson, K.A. Baggerly,
L.R. Ramoth, D.G. Rosen, J. Liu, I. Hellstrom, D. Smith, L. Hartmann, D. Fishman,
A. Berchuck, R. Schmandt, R. Whitaker, D.M. Gershenson, G.B. Mills, R.C. Bast
and N. Carolina, Selection of Potential Markers for Epithelial Ovarian Cancer with
Gene Expression Arrays and Recursive Descent Partition Analysis, Anal. Clin.
Cancer. 10 (2004) 32913200.
36. J.B. Welsh, P.P. Zarrinkar, L.M. Sapinoso, S.G. Kern, C. a Behling, B.J. Monk, D.J.
Lockhart, R. A. Burger, and G.M. Hampton, Analysis of gene expression profiles in
normal and neoplastic ovarian tissue samples identifies candidate molecular
markers of epithelial ovarian cancer, P. Natl. Acad. of Sci. USA. 98 (2001)
11761181.